|
Advertisement |
|
|
||
 |
|
|||||
|
|||||
|
 |
||||||
|
||||||
|
||||||
|
© 2002 American Society for Clinical Oncology Up-to-Date Long-Term Survival Curves of Patients With Cancer by Period AnalysisByFrom the Department of Epidemiology, German Centre for Research on Ageing, Heidelberg, Germany; Finnish Cancer Registry; and Department of Public Health, University of Helsinki, Helsinki, Finland. The work of T.H. was supported by the MaDaMe Project of the Academy of Finland.Address reprint requests to Hermann Brenner, MD, MPH, Department of Epidemiology, German Centre for Research on Ageing, Bergheimer Str 20, D-69115 Heidelberg, Germany; email: brenner@ dzfa.uni-heidelberg.de.
PURPOSE: Provision of up-to-date long-term survival curves is an important task of cancer registries. Traditionally, survival curves have been derived for cohorts of patients diagnosed many years ago. Using data of the Finnish Cancer Registry, we provide an empirical assessment of the use of a new method of survival anlysis, denoted period analysis, for deriving more up-to-date survival curves. PATIENTS AND METHODS: We calculated 10-year relative survival curves actually observed for patients diagnosed with one of the 15 most common forms of cancer in 1983 to 1987, and we compared them with the most up-to-date 10-year relative survival curves that might have been obtained in 1983 to 1987 using either traditional (cohort-wise) or period analysis. We also give the most recent 10-year survival curves obtained by period analysis for the 1993 to 1997 period. RESULTS: For all forms of cancer, period analysis of the 1983 to 1987 data yielded survival curves that were very close to the survival curves later observed for patients who were newly diagnosed in that period (median and maximum difference of 10-year relative survival estimates: 0.9 and 5.7 percent units, respectively). By contrast, the survival curves obtained by traditional (cohort-wise) survival analysis in 1983 to 1987 would have been much lower for most forms of cancer (median and maximum difference: 5.8 and 18.4 percent units, respectively). The 10-year survival curves for the 1993 to 1997 period are substantially more favorable than previously available, traditionally derived survival curves for most forms of cancer. CONCLUSION: Period analysis is a useful tool for deriving up-to-date long-term survival curves of patients with cancer.
MONITORING OF CANCER patient survival is an essential task of both clinical- and population-based cancer registries.1 The most commonly reported measures of patient survival by cancer registries are 5- and 10-year cumulative survival rates,2,3 and the cumulative survival functions (commonly referred to as survival curves) of patients over the first 5 or 10 years or even longer after diagnosis are often graphically displayed.4-6 Traditionally, long-term survival curves have been derived in a cohort-wise fashion, ie, from cohorts of patients diagnosed at least 5 or 10 years ago who have been observed with respect to vital status over 5 or 10 years since then, respectively. Obviously, these survival figures may be outdated at the time they can be derived (ie, after a follow-up of at least 5 or 10 years) in case of recent improvement in prognosis, eg, by advancements in early detection or therapy. To overcome this drawback, a new method of survival analysis, denoted period analysis, was proposed a few years ago.7,8 With this approach, the survival function is exclusively estimated from the survival experience (of patients diagnosed in various calendar years) during some recent time period. However, the method has only been applied by a limited number of cancer registries so far, and the use of period analysis for deriving up-to-date long-term survival curves has not systematically been evaluated empirically. In particular, it is unclear to what extent long-term survival curves derived for some recent time period by period analysis agree with survival curves later observed for patients diagnosed with cancer in that period. In this article, we provide an empirical evaluation of the performance of period analysis for deriving up-to-date long-term survival curves, and we applied this technique to derive recent estimates of 10-year survival curves of cancer patients in Finland.
Patients Our analysis is based on data from the population-based Finnish cancer registry, which is well recognized for its high quality and completeness.9,10 The registry obtains information from many different sources, including hospitals, physicians outside hospitals, and pathologic and cytologic laboratories. Notification of new cancer cases to the registry is mandatory by law, and studies have shown that the registry achieves close to 100% completeness. For example, a linkage with an independent data source, the hospital discharge registry, indicated that in 1985 to 1988 only 965 cases (1.4%) should have been added to the 68,628 cases registered by the Finnish Cancer Registry.10 Mortality follow-up, which is performed by annual record linkage of registry data with death records, is very efficient because of the existence of personal identification numbers. At the time of this analysis, both registration of new cases and mortality follow-up had been complete until the end of 1997. Patients with the 15 most common forms of cancer were included in this analysis, and results are presented separately for each of these forms of cancer.
Methods of Analysis
However, survival curves obtained in that way that pertain to patients diagnosed many years ago might be quite outdated in case of recent improvement in survival. Therefore, we aimed to obtain more up-to-date survival curves by exclusive consideration of the survival experience of patients in a recent time period (for example, the 1993 to 1997 period as indicated by the black squares frame in Fig 1) by period analysis. This analysis includes all patients diagnosed in 1983 to 1997, but the analysis is restricted to the 1993 to 1997 period by left truncation of observations at the beginning of 1993 in addition to right censoring at the end of 1997. With that approach, survival experience during the first year after diagnosis is provided by patients diagnosed between 1992 and 1997, survival experience during the second year after diagnosis is provided by patients diagnosed between 1991 and 1996 and so on, until survival experience during the tenth year of diagnosis, which is provided by patients diagnosed between 1983 and 1988. Of course, it would be of great interest to know how well the 10-year relative curves obtained by period analysis for the 1993 to 1997 period actually predict the 10-year relative survival curves of patients diagnosed in that time period. A definite answer to this question will only be possible 10 years from now. It is possible, however, to evaluate how well period analysis would have predicted meanwhile known 10-year survival curves of patients diagnosed in the past. To evaluate the performance of period analysis for deriving up-to-date survival curves we, therefore, proceeded as follows: we compared the 10-year survival curves actually observed for patients diagnosed in 1983 to 1997 (the most recent cohort of patients who had completed 10-year follow-up by the end of 1997, see Fig 1, solid grey frame) with the most up-to-date estimates of the 10-year survival function that might have been derived at the time of diagnosis of these patients (ie, in 1983 to 1987) using either traditional cohort-wise analysis or period analysis. For simplicity, we neglected any delay in cancer registration, mortality follow-up, and data analysis. In this ideal situation, the traditional cohort estimates available in 1983 to 1987 would have reflected the survival experience in 1973 to 1987 of the cohorts of patients diagnosed in 1973 to 1977, who have completed 10-year follow-up in 1983 to 1987 (see Fig 1, solid black frame). By contrast, the period estimates available in 1983 to 1987 would have exclusively reflected the survival experience in 1983 to 1987 of patients diagnosed between 1973 and 1987 (see Fig 1, dashed black frame). This would have been achieved by inclusion of this whole cohort in the analysis along with left truncation of all observations at the beginning of 1983 (in addition to right censoring at the end of 1987). This type of evaluation allows us to answer the question to what degree the 10-year survival curves later observed for patients diagnosed in 1983 to 1987 could already have been predicted in 1983 to 1987 by the most up-to-date 10-year survival curves available from period analysis or from traditional cohort analysis in that time period. In both cohort-wise and period-wise analyses, a life-table approach was used11 in which conditional survival probabilities were calculated (and combined to derive the cumulative survival function) by intervals of 1 year up to the tenth year after diagnosis. Throughout this article, we present relative, rather than absolute, survival rates.12 Relative survival rates, which are commonly reported by population-based cancer registries, reflect the net mortality due to the cancer. They can be interpreted as the expected survival experience in the hypothetical situation in which the cancer of interest was the only cause of death. Relative survival rates were derived as the ratio of absolute survival rates divided by the expected survival rates of subjects of the corresponding age and sex in the general population, as estimated from population life tables, according to the method commonly known as the Ederer II method13 (with appropriate adaptations for the period analysis approach).
The numbers of patients, the proportion of women, and the median age at diagnosis of patients diagnosed in 1983 to 1987 are shown by cancer site in Table 1. The most common form of cancer in Finland in 1983 to 1987 was lung cancer, followed by breast cancer, colorectal cancer, stomach cancer, and prostate cancer. The vast majority of patients with lung cancer and bladder cancer were men, whereas the proportions of women ranged between 40% and 60% for other nongynecologic forms of cancer. Median age at diagnosis was lowest for patients with cancer of the nervous system (54 years) and highest for patients with prostate cancer (74 years).
Figure 2A-D show the 10-year relative survival curves actually observed for patients diagnosed in 1983 to 1987 (solid grey lines) compared with the most up-to-date estimates of relative survival curves that might have been obtained by period analysis (dashed black lines) and cohort analysis (solid black lines) in 1983 to 1987 (ie, at the time of diagnosis of these patients). The black squares indicate the period estimates of the survival function for the 1993 to 1997 period.
There was tremendous variation in 10-year relative survival rates of patients diagnosed in 1983 to 1987 by cancer site, ranging from approximately 2% for pancreatic cancer to approximately 75% for melanoma. For all forms of cancer, period analysis of the 1983 to 1987 data would have yielded 10-year relative survival curves that were very close (in both shape and levels of survival) to 10-year relative survival curves later observed for patients diagnosed in those years (median and maximum difference of 10-year relative survival estimates: 0.9 and 5.7 percent units, respectively). By contrast, 10-year relative survival curves derived by traditional cohort analysis would have been quite different (mostly much lower) for most forms of cancer (median and maximum difference of 10-year relative survival estimates: 5.8 and 18.4 percent units, respectively). This is because of the fact that there has been substantial improvement in prognosis between patients diagnosed in the 1970s (black solid survival curves) and the 1980s (grey solid survival curves) for most forms of cancer. Obviously, this improvement could not have been captured by the cohort estimates available in 1983 to 1987 because the latter simply reflect the survival experience of patients diagnosed in 1973 to 1977. Among the most common gastrointestinal cancers, improvement in prognosis has been most pronounced for stomach cancer and colon cancer (see grey and black solid survival curves in Fig 2A). For these cancers, the period estimates of 10-year relative survival curves that might have been obtained in 1983 to 1987, in contrast to the cohort estimates, closely match the 10-year relative survival curves later observed for patients diagnosed in this period. Unfortunately, prognosis of pancreatic cancer has not improved and remains very poor. As a result, cohort analysis for patients diagnosed in 1973 to 1977 and 1983 to 1987, as well as period analysis for the 1983 to 1987 period, yield almost identical survival curves. Among the most common gynecologic cancers, improvement has been most pronounced for breast cancer (Fig 2B). However, different from survival curves for other forms of cancer, the slope of survival curves of patients with breast cancer hardly levels off during the first 10 years after diagnosis. This pattern is the same for both period and cohort estimates of survival curves, but the former are again closer to the survival curves actually observed for newly diagnosed patients in 1983 to 1987, although agreement is less pronounced than for other forms of cancer, such as cancer of the ovaries. A different picture is seen for invasive cancer of the uterine cervix. Ten-year relative survival of patients diagnosed with invasive uterine cervix cancer in 1983 to 1987 is lower than survival of patients diagnosed in 1973 to 1977. This finding can be explained by the effects of widespread screening for precursors of cervical cancer by Pap smear, which led to a strong decrease in incidence of invasive cervical cancer over time (by detection and removal of precancerous lesions) along with a shift of the age distribution of patients with invasive cervical cancer to women at older ages, who tend to have much poorer survival.3,14 It is also more likely that the precancerous lesions picked by the screening would have developed slower as invasive tumors than those not picked by screening leading, in the case of Pap screening, to an overrepresentation of unfavorable stages among the invasive tumors.15 As expected from theory,7,8 the reduced prognosis of patients diagnosed in the 1980s is also more closely approximated by survival curves generated by period analysis than by survival curves generated by cohort analysis. Overall, survival curves are much more favorable for cancer of the corpus uteri, but there has been little change over time, and similar survival curves would have been obtained by period and cohort analysis. Survival curves have strongly improved over time for the two most common forms of cancer of the urinary tract, cancer of the kidneys and the urinary bladder, and somewhat less so for cancer of the prostate (Fig 2C). The slope of survival curves of prostate cancer also remains quite steep over time, reflecting a relatively high proportion of late deaths for this form of cancer. Regardless of such variation in the shape of survival curves, estimates of survival curves that would have been obtained in 1983 to 1987 by period analysis are again much closer to the survival curves later observed for patients diagnosed in that period than the corresponding estimates that might have been obtained by cohort analysis. The same patterns also hold for other common forms of cancer shown in Fig 2D; survival curves have strongly improved over time for these cancers (melanoma, cancer of the nervous system, and leukemia) despite their different levels of long-term survival rates. Unfortunately, prognosis continued to be very poor for lung cancer, which continues to be a common form of cancer in Finland. Accordingly, all types of survival analysis yielded almost identical survival curves for this type of cancer. We would also like to address the question of more recent 10-year relative survival curves. At the time of this analysis, registry data and mortality follow-up were complete until the end of 1997, and it will take another 10 years until we can derive the 10-year relative survival curves of patients diagnosed in the mid-1990s. However, from our empirical evaluation, we feel confident that period estimates for the 1993 to 1997 period, which can be derived from existing data, do come quite close to the 10-year relative survival curves of patients diagnosed in that period. The period estimates for 1993 to 1997 are depicted by black squares in Fig 2A to 2D. They suggest that there has been quite substantial further recent improvement in survival curves for many common forms of cancer, which would so far remain undisclosed in the most up-to-date survival curves obtained from cohort estimates (ie, the survival curves for the 1983 to 1987 cohorts, the grey lines in Fig 2A to 2D). In particular, recent 10-year survival curves are now much more favorable for all common gastrointestinal cancers except pancreatic cancer, all common gynecologic cancers, and leukemia, but the most dramatic improvement is disclosed for cancers of the urinary organs, most notably kidney cancer.
Our results give further empirical confirmation to the suggestion that period-wise analysis of survival will provide more up-to-date estimates of long-term survival rates than traditional cohort-wise analysis of survival (unless survival rates do not change over time, in which case period-wise and cohort-wise analyses yield identical results).7,8 Our article illustrates that period analysis also preserves the characteristic shape of long-term survival curves of cancer patients. This suggests that survival curves produced by period analysis are a useful supplement to the tools of cancer surveillance, which should be as up-to-date as possible. So far, there have been few applications of period analysis to produce up-to-date long-term survival curves.16-18 Apart from the fact that the method has been introduced only a few years ago, this may have been a result of the lack of empirical evaluation (of the kind presented in this article). Further obstacles against the use of period analysis may have been the lack of pertinent computer programs, as the method is neither included in standard commercial software packages nor in special computer programs for survival analysis that are widely used by cancer registries.19,20 To overcome this obstacle, we have recently developed a user-friendly computer program for period analysis of both absolute and relative survival rates that we will be glad to share with colleagues in the field who are interested in its use. Whereas the period approach has not traditionally been used in survival analysis, it is well established in related fields. For example, the period approach is probably the most commonly applied approach to estimate population life tables, and it is almost universally used to derive up-to-date estimates of life expectancy. Population life tables provide special types of long-term survival curves, ie, the survival curves of populations of newborns over a lifespan.21 Nevertheless, up-to-date estimates of these survival curves are usually not derived for real cohorts of newborns (which could only be done a lifespan after their birth) in a cohort-wise fashion. Rather, various parts of these survival curves are typically estimated in a period-wise fashion from the survival experience of various birth cohorts in some recent period (eg, a recent calendar year). A commonly used measure in the field of cancer epidemiology is the cumulative risk to develop a certain cancer up to a certain age (eg, age 75) in the absence of competing causes of death.22 The complementary measure, 1-CR, is the cumulative probability of cancer-free survival up to a certain age in the absence of competing causes of death. Again, this special measure of cumulative survival is not estimated in a cohort-wise fashion from real cohorts of newborns (which could only be estimated with long delay, eg, after 75 years). Rather, it is estimated in a period-wise fashion from cancer incidence at various ages within a recent time period. The following limitations of our analyses should be kept in mind: although we presented survival curves for each of the 15 most common forms of cancer, space limits did not allow to present results of more detailed analyses. For example, no stratification by age and stage at diagnosis was made. Therefore, the survival curves presented in this article do not, by themselves, allow any conclusions as to the reasons for improvement (which was observed for most forms of cancer) or deterioration (which was transiently observed for cervical cancer) of prognosis over time. For example, observed increases in survival rates over time may have been a result of earlier diagnosis or even overdiagnosis, advancements in therapy, or both, and the importance of these reasons is likely to have strongly varied between different forms of cancers. More detailed analyses (eg, taking age and stage at diagnosis into account), which could be performed in a period-wise manner in the same way as for the overall survival rates shown in this article, are needed to disentangle those reasons. Notwithstanding the need for more detailed analyses to elucidate reasons for changes in survival curves over time, the very consistent patterns disclosed in our analyses should encourage expanding the use of the period approach, the benefits of which have long been accepted and appreciated in other applications, to the monitoring of cancer patient survival. For those cancers with recent improvement in prognosis, the more up-to-date survival curves obtained by period analysis of survival may help to prevent both patients and their clinicians from being burdened by unduly pessimistic survival expectations, which are suggested by previously available survival estimates.
1. Parkin DM, Hakulinen T: Analysis of survival, in: Jensen OM, Parkin DM, Maclennan R,et al (eds): Cancer Registration: Principles and Methods. Lyon, France, International Agency for Research on Cancer, IARC Scientific Publications no. 95, 1991, pp 159-176 2. Berrino F, Capocaccia R, Estève J, et al (eds): Survival of Cancer Patients in Europe: The EUROCARE-2 Study. Lyon, France, International Agency for Research on Cancer, IARC Scientific Publications no. 151, 1999 3. Dickman PW, Hakulinen T, Luostarinen T, et al: Survival of cancer patients in Finland 1955-1994. Acta Oncol 38: 1-103, 1999 (suppl 12)
4.
Adami H-O, Sparén P, Bergström R, et al: Increasing survival trend after cancer diagnosis in Sweden: 1960-1984. J Natl Cancer Inst 81: 1640-1647, 1989
5.
Nab HW, Hop WCJ, Crommelin MA, et al: Changes in long term prognosis for breast cancer in a Dutch cancer registry. BMJ 309: 83-86, 1994 6. Wingo PA, Gloeckler Ries LA, Parker SL, et al: Long-term cancer patient survival in the United States. Cancer Epidemiol Biomark Prev 7: 271-282, 1998[Abstract] 7. Brenner H, Gefeller O: An alternative approach to monitoring cancer patient survival. Cancer 78: 2004-2010, 1996[CrossRef][Medline] 8. Brenner H, Gefeller O: Deriving more up-to-date estimates of long term patient survival. J Clin Epidemiol 50: 211-216, 1997[CrossRef][Medline] 9. Parkin DM, Muir CS, Whelan SL, et al (eds): Cancer Incidence in Five Continents. Volume VII. Lyon, France, International Agency for Research on Cancer, IARC Scientific Publications no. 120, 1997 10. Teppo L, Pukkala E, Lehtonen M: Data quality and quality control of a population-based cancer registry: Experience in Finland. Acta Oncol 33: 365-369, 1994[Medline] 11. Cutler SJ, Ederer F: Maximum utilization of the life table method in analysing survival. J Chron Dis 8: 699-712, 1958[CrossRef][Medline] 12. Ederer F, Axtell LM, Cutler SJ: The relative survival rate: A statistical methodology. Natl Cancer Inst Monogr 6: 101-121, 1961 13. Ederer F, Heise H: Instructions to IBM 650 programmers in processing survival computations. Methodological note No.10. End Results Evaluation Section. Bethesda, MD, National Cancer Institute, 1959 14. Gustafsson L, Ponten J, Zack M, et al: International incidence rates of invasive cervical cancer after introduction of cytological screening. Cancer Causes Control 8: 755-763, 1997[CrossRef][Medline] 15. Saxén E, Hakama M: Cancer illness in Finland with a note on the effects of age adjustment and early diagnosis. Ann Med Exp Biol Fenn 42: 1-28, 1964 (suppl 2) 16. Brenner H, Stegmaier C, Ziegler H: Recent improvement in survival of breast cancer patients in Saarland, Germany. Br J Cancer 78: 694-697, 1998[Medline] 17. Brenner H, Stegmaier C, Ziegler H: Trends in survival of patients with ovarian cancer in Saarland, Germany, 1976-1995. J Cancer Res Clin Oncol 125: 109-113, 1999[CrossRef][Medline] 18. Brenner H, Kaatsch P, Burkhardt-Hammer T, et al: Long term survival of children with leukemia achieved by the end of the 2nd millenium. Cancer 92: 1977-1983, 2001[CrossRef][Medline] 19. Hakulinen T, Abeywickrama K: A computer program package for relative survival analysis. Comp Prog Biomed 19: 197-207, 1985 20. Voutilainen ET, Dickman PW, Hakulinen T: SURV2: Relative Survival Analysis Program, version 2.02ß. Helsinki, Finland, Finnish Cancer Registry, 1998 21. Estève J, Benhamou E, Raymond L: Statistical Methods in Cancer Research, Volume IV. Descriptive Epidemiology. Lyon, France, International Agency for Research on Cancer, IARC Scientific Publications no. 128, 1994 22. Breslow N, Day NE: Statistical Methods in Cancer Research, Volume II. The Design and Analysis of Cohort Studies. Lyon, France, International Agency for Research on Cancer, IARC Scientific Publications no. 82, 1987 Submitted May 29, 2001; accepted October 2, 2001.
This article has been cited by other articles:
|
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
|
|||||||||||
|
|||||||||||
|
|
Copyright © 2002 by the American Society of Clinical Oncology, Online ISSN: 1527-7755. Print ISSN: 0732-183X
|
TOP
ABSTRACT
INTRODUCTION
